Hugging Face Overview
T-lite-it-2.1: An Advanced Russian LLM with Tool-Calling
T-lite-it-2.1 is an 8 billion parameter model from t-tech, based on the Qwen 3 architecture, specifically engineered for the Russian language. It introduces significant advancements in instruction following and, crucially, adds robust tool-calling capabilities, a key differentiator from its predecessor, T-lite-it-1.0. This model is optimized for higher Russian text generation throughput, making it efficient for various applications.
Key Capabilities
- Enhanced Tool-Calling: Outperforms Qwen3-8B in tool-calling scenarios, making it highly suitable for agentic applications and complex workflows.
- Superior Instruction Following: Demonstrates strong performance in adhering to diverse and complex instructions.
- Russian Language Optimization: Built with an optimized tokenizer for efficient and high-throughput Russian text generation.
- Expert Merging Strategy: Utilizes a unique SLERP-based expert merging approach, combining specialized 'IF Expert', 'General Expert', and 'Tool-Call Expert' for balanced and superior performance across different domains.
- Long Context Support: Natively supports a context length of 32,768 tokens, with recommendations for extending it further using
rope_scaling.
Good For
- Agentic Applications: Ideal for scenarios requiring the model to interact with external tools and APIs.
- Complex Russian Workflows: Excels in tasks demanding precise instruction following and multi-step tool use in Russian.
- General Russian Chatbots: Capable of handling a wide range of general and conversational tasks in Russian.
- Benchmarking: Achieves strong results across various benchmarks, including Ru Arena Hard (83.9), ruIFeval (75.9), and ACEBench (61.0), demonstrating significant improvements over T-lite-it-1.0 and competitive performance against other 8B models.